Requirements Writing

Overview

A requirement is a promise about behavior, written so QA can later prove it and a developer can build it without guessing. The recurring failures are requirements that specify implementation (HOW), bundle several behaviors into one, or hide assumptions that later surface as defects.

Core principle: State WHAT and WHY. HOW belongs to the architect and developers.

The shape

• User story: As [persona], I want [capability], so that [outcome].
• Context: why now, what problem it solves.
• Constraints: hard limits (locale, performance, compliance) treated as given.
• Assumptions: stated explicitly — an unstated assumption becomes a defect.

Rules

• WHAT/WHY, not HOW — but field-aware. description and acceptance_criteria state WHAT and WHY: "Users can export tasks to CSV" — not "add a /export endpoint using library X." Endpoints, files and tables belong only in technical_description, only once verified with code tools (get_repo_tree, codebase_search/grep_code), and only for small direct tasks you write yourself. A criterion may name an endpoint only when the API itself is the product surface (see acceptance-criteria-gwt's API-surface example) — not as a stand-in for what a user sees.
• Independently testable. If a requirement contains "and", consider splitting.
• No solutioning in description/acceptance_criteria. Naming a datastore, framework, or endpoint there is an architect/dev decision leaking in, even though the same name is correct in technical_description.

Worked Example

❌ "Add a Redis cache to the task list endpoint so it's fast."
   (HOW — names the tech; "fast" is untestable)

✅ Story: As a project member, I want the task list to load quickly,
          so that I can scan my board without waiting.
   Context: boards with 500+ tasks feel sluggish today.
   Constraint: list renders < 500ms at p95 for 1000 tasks.
   Assumption: 1000 tasks is the realistic upper bound per project.

The architect chooses caching vs indexing vs pagination — the requirement fixes only the outcome and the bound.

Common Mistakes

• Prescribing the implementation.
• "Fast/intuitive/robust" with no measurable target.
• One requirement covering three behaviors.
• Assumptions left unstated.

Red Flags

• The requirement names a library, table, or endpoint.
• You can't say how QA would prove it.
• An adjective where a number belongs.